英国 AI 安全研究所：OpenAI与 Anthropic 旗舰模型测试中现"未授权"攻击行为

英国 AI 安全研究所：OpenAI与 Anthropic 旗舰模型测试中现"未授权"攻击行为

2026-08-05 09:12

ChainThink 消息，8 月 5 日，据彭博社报道，英国政府 AI 安全研究所周二披露，在对 OpenAI 的 GPT-5.6-Sol 和 Anthropic 的 Mythos 5 进行安全评估时，两款旗舰模型均出现“未经授权”的有害行为。

相关行为包括主动入侵真实网站、尝试向软件注入恶意代码，且目标涉及真实人群与组织。测试期间，研究所为模型开放互联网访问权限，并关闭部分安全过滤器，以评估其极限能力。

该研究所成立于 2023 年。

Disclaimer: Contains third-party opinions, does not constitute financial advice

Recommended Reading

NVIDIA attracts $85 billion in investor demand during massive bond issuance

06-16
NVIDIA attracts $85 billion in investor demand during massive bond issuance

Ethereum surges over 10% in 24 hours, currently priced at $1,841.31

06-16
Ethereum surges over 10% in 24 hours, currently priced at $1,841.31

Amazon announces a multi-billion dollar investment in Missouri to build a data center campus, expected to create over 400 long-term positions

06-16
Amazon announces a multi-billion dollar investment in Missouri to build a data center campus, expected to create over 400 long-term positions

Binance Platform's SpaceX Perpetual Contract Trading Volume Surpasses $9 Billion, Capturing Over 60% Market Share

06-16
Binance Platform's SpaceX Perpetual Contract Trading Volume Surpasses $9 Billion, Capturing Over 60% Market Share

Binance platform XLM/USDT short-term spike down to $0.17, now recovered to $0.225

06-16
Binance platform XLM/USDT short-term spike down to $0.17, now recovered to $0.225

Trump: The Strait of Hormuz has been fully reopened as of Friday, and all agreements have been signed

06-16
Trump: The Strait of Hormuz has been fully reopened as of Friday, and all agreements have been signed

SlowMist: Aztec Connect Contract Hacked for $2.19 Million Due to ZK-Rollup L1/L2 State Boundary Vulnerability

06-15
SlowMist: Aztec Connect Contract Hacked for $2.19 Million Due to ZK-Rollup L1/L2 State Boundary Vulnerability